DiLu: A Knowledge-Driven Approach to Autonomous Driving with Large Language Models
Authors: Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao MA, Pinlong Cai, Min Dou, Botian Shi, Liang He, Yu Qiao
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments prove Di Lu s capability to accumulate experience and demonstrate a significant advantage in generalization ability over reinforcement learning-based methods. Moreover, Di Lu is able to directly acquire experiences from real-world datasets which highlights its potential to be deployed on practical autonomous driving systems. |
| Researcher Affiliation | Collaboration | 1 Shanghai Artificial Intelligence Laboratory 2 East China Normal University, 3 The Chinese University of Hong Kong |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Project page: https://pjlab-adg.github.io/Di Lu/ |
| Open Datasets | Yes | In our experimental setup, we utilize the well-established Highway-env as our simulation environment, which is a widely used platform in the fields of autonomous driving and tactical decision-making (Leurent, 2018). ... To illustrate this capability, we created two Memory modules, each containing 20 experiences extracted from two distinct sources: (1) Highway-env and (2) City Sim, a dataset comprising real-world vehicle trajectory data (Zheng et al., 2023). |
| Dataset Splits | No | The paper mentions conducting experiments but does not explicitly provide specific dataset split information (e.g., exact percentages or sample counts for training, validation, and test sets, or specific citations to predefined splits with authors and year for reproducibility). |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | Yes | Within our Di Lu framework, the large language model we adopt is the GPT family developed by Open AI. GPT-3.5 (Open AI, 2023) is used in the Reasoning module of the framework, which is responsible for making reasonable decisions for the ego vehicle. GPT-4 is used in the Reflection module since it has demonstrated significantly improved self-repair and fact-checking capabilities compared to GPT-3.5 (Bubeck et al., 2023; Olausson et al., 2023). To serve as the Memory module in the Di Lu framework, we adopt Chroma1, an open-source embedding vector database. The scenario descriptions are transformed into vectors using the text-embedding-ada-002 model of Open AI. |
| Experiment Setup | Yes | In terms of the setup for highway-env, we directly obtain vehicle information from the underlying simulation and input it into the scenario descriptor. This information only includes each vehicle s position, speed, and acceleration data in the current frame, without any decision intent or potential risk information, as shown in Figure 7. Meta-actions are used as the decision output in our experiments, which include five discrete actions to control the ego vehicle: acceleration, maintaining speed, deceleration, and lane changes to the left and right. For each closed-loop driving task, we define a successful completion time of 30 seconds, with a decision-making frequency of 1Hz. This means that if the ego vehicle can navigate through traffic at a reasonable speed and without collisions for 30 seconds, we consider the task to be successfully completed. Unless otherwise stated, our experimental environment is a four-lane motorway with a vehicle density of 2.0, representing scenarios with relatively high traffic density and complexity. All other settings follow the simulator s default configurations. |